| # Integration acceptance criteria |
|
|
| Use this short list to decide whether the primitive fits your existing service. |
|
|
| 1. Express the statistic as `(N0 + sum(w*v))/(D0 + sum(w))`, with nonnegative |
| bounded weights and a strictly positive denominator floor. |
| 2. Establish valid input intervals. If your feasible set is correlated, a |
| surrounding independent box gives sufficient positive guarantees; its |
| negative witnesses may be relaxation artifacts. |
| 3. Choose `exact` for the unconditional integer-input numerical contract, or |
| choose `filtered` explicitly under the published IEEE/NumPy premises. |
| 4. Prepare static values once. For the supported fast use case, submit fresh |
| mass arrays and one threshold vector; keep the running process alive. |
| 5. Handle false results as coordinatewise counterexample existence. Ask for |
| an exact witness only when the application can use it. Do not combine |
| separate coordinate witnesses into a simultaneous state. |
| 6. Compare total service cost on representative inputs, including JSON/array |
| conversion, construction, fallback and any requested witnesses. Use cached |
| exact bounds for many policies on the same unchanged mass box. |
| 7. Run the independent tests and near-boundary fixtures on the deployment |
| environment. Keep the exact backend available as a reference. |
|
|
| No model training, GPU, dynamic tuning or threshold tolerance is necessary. |
| The application threshold is the policy itself; the numerical allowance is |
| computed from arithmetic precision and row count rather than learned. |
|
|
